feat: keyword-based message routing for pre-turn model selection

[scottgl9]

Summary

Add zero-cost pre-turn model routing that selects the LLM based on keyword/pattern matching in the incoming message, before any LLM call.

Motivation

Different tasks benefit from different models (coding → claude-sonnet, research → gemini-flash, quick Q&A → qwen35). Currently requires manual model switching.

Proposed Design

messageRouting: {
  rules: [
    { match: ["code review", "PR"], model: "github-copilot/claude-sonnet-4.6" },
    { match: ["research", "search"], model: "gemini-3-flash" },
  ],
  default: "qwen35",
}

• First-match-wins rule ordering
• Case-insensitive substring matching
• Zero LLM cost — pure string matching
• Falls back to primary model if no rule matches

Implementation

Prototype: scottgl9/nanoclaw feature/message-routing-20260406

• src/message-routing.ts — types + resolveMessageRoutingModel()
• src/message-routing.test.ts — 6 tests passing
• src/types.ts — messageRouting field added
• src/index.ts — wired into agent dispatch

Also implemented in openclaw/openclaw#62153 and scottgl9/leanclaw#7.